keras - > Tensorflow.js ->(html + css + javascript)->github pages
To make a convolution neural network to recognise handwritten digits by training the model on MNIST dataset available in keras.
The training dataset contain 60000 images and testing contain 10000 images. Each image is 28x28 pixel and grey scale.
⚈ It is a 17 layer model with Conv2D,MaxPooling2D,BatchNormalization,Dense,Flatten and Dropout layer combination.
⚈ Input layer has 32 neuron and output layer has 10 neurons as 10 different clases exsist.
⚈ 30 epochs are used.
⚈ Categorical_loss is loss function and adam is used for optimization.
⚈ Model gives 99.15% accuracy.
Save model using tensorflowjs converters as json file and weight as .h5 file.Use Tensorflow.js to load model and predict in javascript file.